PACKAGING TECHNOLOGY
Electronic systems are usually implemented as hierarchical packages of components. Passive or active electronic elements, such as resistors and transistors, and their wiring are typically combined into memory or logic units, which are then combined into circuits and devices, which are combined into larger functional units, and so forth up to the level of a system.
Each higher level of hierarchy grants the designer greater productivity, but compounds costs for connecting packages together logically and physically. Communication of data and timing among devices at each of these levels requires signal interconnection means, which the package provides. The package also provides powering means and fulfills other requirements such as physical support, heat removal and protection.
By convention, there are five hierarchical packaging levels 0-4, although these levels are not rigidly defined. An exemplary scale in the hardware system hierarchy is the naked semiconductor wafer, sometimes called an undiced "Level 0" package. Many components are formed simultaneously on a common substrate during fabrication stages, even if the substrate is subsequently separated into subunits. For instance, wiring, memory or logic gates may be assembled into integrated circuits on the surface of a semiconductor wafer and then cut into individual dies. Examples of dies include microelectronic devices implemented on semiconductor materials, superconductors bearing Josephson Junctions, and materials bearing other quantum interference devices.
Individual dies are typically mounted into "Level 1" packages, which provide mechanical stability, protection, cooling and heat dissipation, power and grounding, and interconnection of signal lines (including clocking) to other packages. Examples include DIP, ceramic, surface mounted and socketed packages.
A "Level 2" package is a module carrying one or more Level 1 or Level 0 packages and interconnecting their signal and power wiring. It typically comprises a printed circuit board (PCB), a printed wiring board (PWB), or a thermal conduction module, and may cluster one or more interconnected packages for these purposes. "Level 3" assemblies further organize the Level 1 and Level 2 packages, typically with backplanes, but do not differ conceptually from Level 2 or Level 1 packages. The "Level 4" package canonically ties together the lower level packages with power supplies, environmental systems, mechanical systems, peripherals, and so forth to provide system functionality.
A multichip module ("MCM" or "Level 1 1/2" package) provides modular functionality as a Level 2-like or Level 3-like package for holding and interconnecting multiple dies and/or associated interconnections. At a minimum, an MCM provides the signal distribution, and power is usually distributed by way of the MCM as well. The MCM may also, or merely, encapsulate its constituent dies as an erstwhile Level 1 or Level 0 package, thereby providing protection. It may also communicate the dies to a heat sinking substrate, thereby providing heat dissipation.
Strictly speaking, an MCM could be treated as a package at any level of hierarchy, as defined in practice by its interconnection topology. Note that the term "chip" is used interchangeably in the industry both in reference to Level 0 dies and Level 1 packages (e.g., a multi chip module is usually in fact a multidie module). As used herein, the term chip refers to a Level 0 package or die unless context indicates otherwise. The term "module" or "submodule" as used herein is intended to be general, and can refer to any package level, for example one or more Level 0 dies, one or more Level 1 or Level 0 chips (packaged or not), and of course higher order ensembles.
PRESENT MCM TECHNOLOGIES
An MCM involves two or more dies, whether bare or encapsulated, mounted and conductively coupled to it. It provides power and inter-die signal wiring. In some MCM technologies, the dies are physically bonded to a substrate, and leads that are wire-bonded to peripherally positioned contacts (e.g., pins) supply the conductive connections between the dies and a multichip substrate. Other technologies utilize a "flip-chip" configuration in which the leads of the dies are positioned either peripherally or over much of the die area (such as a pin grid array (PGA) or solder bumps) and are soldered or otherwise bonded to respective contacts on the multichip substrate.
Several families of multichip packaging technology are standard at present. The so-called MCM-L technology utilizes a laminated, organic board substrate to which dies are bonded by flip-chip, tape automatic bonding (TAB), or wire-bonding. In the MCM-C technologies, dies are attached either directly by flip-chip or indirectly in prepackaged carriers to a ceramic thick-filmed substrate. The ceramic substrate is formed either sequentially, by a print and fire process, or by lamination and sequential co-insertion of screened green sheets. The MCM-D technology utilizes deposited thin-film substrates to which dies are then attached as in MCM-C. There are also variations of these three basic MCM technologies. For example, a variation using plastic packages and involving molding compounds and lead frames is the so-called multichip plastic quad packs (MCM-P) technology. Another recent variation involves the use of deposited thin films on a ceramic multichip substrate, referred to alternatively as MCM-DC or MCM-CD, and typically provides inter-die signal wiring in the deposited polymer-metal thin-film layers and power/ground wiring in the co-fired ceramic thick-filmed substrate. All of these technologies are subjects of intense research and invention in industry and universities. (See Rao R. Tummala, "Multichip Packaging--A Tutorial."Proc. of the IEEE, December, 1992.)
Many approaches to the construction of high density multichip modules have been proposed. The IBM C4 technology attaches dies to the multichip module in a flip-chip face-down configuration. The arrangement minimizes the parasitic inductance of the package leads, and allows pad location at any point on the interior of the die. Typically, dies are attached to the module using a reflow-solder approach. Dies are bumped by bonding to each pad several layers of protective metalization followed by a 10-200 micron diameter solder ball. A plurality of dies are then accurately positioned on the multichip module, and reflow-soldered into place. Inspection of the solder joints can be done with thermographic or radiographic techniques, but may be difficult otherwise. Pad location is no longer limited to the die periphery, but is often constrained by thermal coefficient of expansion mismatch between silicon and module to lie within some radius of the center of the die. Repair is carried out by module heating, die removal, and reflow-soldering of replacement die.
"Chips-first" face-up wire-bonding of silicon dies to high density silicon, ceramic, or copper-polyimide modules is similar to conventional hybrid manufacturing technologies, and shares difficulties in rework and bonding yield. The GE/TI process forms a planar wafer-like module from collections of selected loose dies placed face down on a flat surface, and then encapsulates it in a polyimide carrier. After curing, this carrier is flipped over, planarized, and used as a module for further (possibly multilayer) metalization. The major advantages of metalizing on top of planar ensembles of dies include the fine lithography achievable and very small interconnect parasities.
The process can also be generalized to stacks of dies in 3D. In the Irvine Sensors approach, dies with electrically insulated backs are designed with leads fanned-out to contact pads lined up along an edge, and stacked with those side pads aligned precisely. The exposed side is then cleaned and polished, and interconnect lines are metalized onto it. The package is extremely dense, although heat dissipation can become limiting, but leads must still travel out to the edge and back to route to other die.
In the Cray Computer approach, holes are drilled into a stack of dies in a grid, plated to form a conductive contact with die vias, and then threaded with twisted gold wires, separating the die slightly and supporting them. The package is extremely dense and conceptually trivial, but extremely difficult to manufacture.
NEED FOR MCM TECHNOLOGY
In the present art, it costs much more (in terms of power, latency, performance and circuit real estate) to signal off-die than to stay on-die. Off-die signalling designates communicating between two points, on the same die or different dies, using off-die wiring. A major design goal with the present art is therefore to increase the number of circuits on each die, increasing the ratio to expensive off-die interconnections. However, as the size of a die approaches some economically and technologically feasible limit, the probability that randomly occurring manufacturing defects will produce an unacceptable die rises exponentially in a Poisson distribution. Since slightly larger die yield at significantly lower rates, this so-called "surface-to-volume" (communication-to-computation) ratio severely constrains fabrication yield, hence cost per functioning die.
The need for multichip technologies arises in large part from this inability to produce arbitrarily large semiconductor dies with acceptable yield. Practical limits on die size for a given technology force system designers to partition large digital systems among multiple dies. Unfortunately, such partitioning dramatically impacts system performance since inter-die communication typically inflates packaging costs by tens to many hundreds of percent.
Multichip technologies also offer the possibility of reducing the cost of intra-die communication in modern CMOS technologies. If off-die signalling became somewhat less expensive than at present, at some point it would become preferable to on-die wiring for certain intra-die communication. The implications of this need to be recognized by practitioners of today's MCM art: Dies should be made smaller, at the expense of more interconnects per noncommunication gate. Smaller dies have significantly higher yield and lower cost per area, so if designers could employ cheap inter-die communication on MCMs, they could transparently treat a multichip module as if it were a single enormous ensemble of electronic devices (including their interconnections), using small, very high yielding monolithic integrated circuit dies (and modules of them) as the subunits. This would be tremendously useful for designers and cost-effective for semiconductor foundries.
Use of MCMs for intra-die signalling should also have important attractions for low impedance--hence low dispersion--signalling between points on the same die. Low dispersion broadcast is essential for clock distribution, and low dispersion point-to-point is useful in general.
Another application for multichip technologies is in customizing on-die interconnects or engineering change pads. It is valuable to customize interconnects in the manner of a breadboard, within dies up to full wafer-size or among ensembles of dies, in order to rig together specialized functions, test performance, edit out defects, redefine a system's basic connectivity, or perform the functions of engineering change pads. At present, customization generally requires a lengthy (circa six week) logic array masking process, unreliable laser fusing/breaking of wiring junctions, significant expense in microfabrication of wiring, macro-scale (e.g., wire-wrap) assemblages, lack of durability (e.g., hand-wired jury-rigs), or a combination of these. The enormous value of customization can be exploited with properly designed MCMs.
Inter-die signalling technology trends demand MCM technology. Shorter signal paths directly contribute to higher performance, since they allow shorter delays and faster clock rates (i.e. more operations per second) and wider synchronous instructions (i.e. more operations per clock tick). Dies are generally planar, so the longest signal path will scale as the diameter, or roughly as the square-root of the area. Planar MCMs can be constructed more densely with smaller dies since they ask less wiring overhead for excursions to the edge and back. In principle, the electronic devices a die bears that make it useful could be disposed volumetrically, but in the current art that is uncommon, except for on-die wiring and stacked capacitors. MCMs can in principle be constructed in a space-filling manner, so that the volumetric packing is denser than planar packing, hence signal paths are kept much shorter (e.g., proportional to the cube root of the chip package volume, which is small if the package is thin compared to its area).
The cost of interconnection networks for signal lines generally scales as a function of area (i.e. a proxy for system real estate) and number of layers (i.e. density and layout efficiency), with linear scaling as a goal. The number of leads or number of chips within an area may inflate cost by a significant factor in many packaging technologies, such as those using wire-bonds. Even if a technology avoids scaling as the number of leads or die, if it requires post-processing steps to form a package, as with solder bump conductive couples, it may still be expensive due to yield losses from handling, amortization of costly test/repair cycles, and of course operating and capital costs.
Even if a large and a small system have identical numbers of wires starting and stopping in each average square inch, the wires in the larger system will travel further on average, so account for a greater proportion of system density than in the smaller system. This is the well-known "Law of Numbers," and is a dominant consideration in constraining layout to a minimum number of metalized layers, which generally rises (geometrically) faster than real estate as a chip design grows. Runaway wiring density can in principle choke off the manufacturability of large high performance systems. The requirement for high interconnection density makes MCMs almost inevitable, and rewards volumetric packing as well as the shorter Manhattan distances that accrue to volumetric interconnection.
A further need for multi-chip technology arises from the cost and complexity of combining hybrid materials to exploit properties of each. For instance, an arbitrarily large, cheap silicon CMOS chip would still lack the speed and optical properties of GaAs, while growing one material on the other is inherently more complicated than forming them separately.
A further need arises from the enduring value of packaging hierarchies, wherein standardized packages with well-described components, such as microprocessors, are available commercially with various advantages compared to components or full system-level packaging. Important advantages to using such hierarchies may include lower cost, modular upgradeability, well-characterized behavior, and multiple sourcing. The designer balances competing benefits and costs of a hierarchy: necessarily limiting the degrees of freedom of system design, while commensurately reducing complex system interactions and failure modes.
A further need arises from the difficulty of package manufacture per se. The engineering or manufacturing complexity, process requirements, and cost of the package approach or exceed those of the die, so the cost and turnaround time for the package can become as formidable as those of the die.
PROBLEMS OF PRESENT MCM TECHNOLOGIES
Despite the intensive research efforts of the past several years, present day MCM technologies still have significant problems in terms of cost, performance, design, manufacturability, reliability and reparability, as well as shortcomings with respect to the needs enumerated above.
Present MCM technologies require significant retooling and/or expensive reorientation of existing integrated circuit ("IC") fabrication lines. High volume is needed to realize cost advantages of MCM packaging, but re-implementing an existing production system (for which high volume demand already exists) to utilize MCM technology typically requires extensive system-wide redesign. System vendors rationally resist such efforts, choosing instead to implement certain clusters of ICs as application specific Ics (ASICs), which generally involves only local redesign of Level 2 boards and Level 1 chips. Accordingly, the relatively high up-front cost of MCM implementation discourages use of MCM technology for systems where large volume cannot be predicted a priori. (See Balde, J.W., "Crisis in Technology: The Questionable U.S. Ability to Manufacture Thin-Film Multichip Modules." Proc. of the IEEE, December, 1992.)
While the electrical performance of TAB, fine-line, or solder bump conductive interconnect can be significantly better than wire-bonding for a conductive interconnect, and pad count is somewhat less constrained, prior art MCM processes often require special processing of wafers or dies prior to assembly to place solder balls or construct metal bonding locations for assembly. The processes also require custom tooling and substantially more sophisticated wafer manufacture post-processing than would be required for standard wafer or packaging lines.
Reparability and die attach yield issues also arise with current MCM technology, principally because of the difficulty in die removal and replacement. Testing the dies in the MCM before it is fully assembled (and paid for) typically accounts for tens of percent of the delivered MCM cost, due to the need for sacrificial test rigs or time-consuming intermediary connections as well as the cost of compensating for parasities in order to test at operating speeds. Making physical contact with microscopic probes or rigs of probes is slow, and exposes the probe points to mechanical forces leading to misalignments, fatigue, and wearing. If the dies are packaged beyond Level 0, only a fraction of the interconnects may even be visible. Nevertheless, working dies should be selected perfectly before module assembly, due to the expense (or impossibility) of reworking the module. Repair methods are very difficult--essentially a tear down and rebuild--and usually require essentially the same elaborate, expensive assembly technology as used to build the MCM originally.
The difficulties, mechanical constraints and costs of multichip module designs are largely driven by the drawbacks of conductive signalling per Conductive signalling is almost universally understood to be inevitable among practitioners of the present art (See Daryl Ann Doane and Paul D. Franzon, Multichip Module Packaging Technology and Alternatives. Van Nostrand Reinhold, 1993). Conductors in general must expose a face in contact to one another; they cannot hide behind shielding or a passivation layer. If there are (re)movable constituents in the electrical path, there will be exposed surfaces, and unless certain materials (e.g., noble metals) are used on exposed surfaces, conductors may be susceptible to oxidation, although some non-metallic conductors avoid the oxidation problems. Troublesome surface chemistry complications include mechanical stability (e.g., whiskering), finite conductance (e.g., charge carrier saturation), and time-dependent material or phase changes (e.g., intermetallic compounds). Non-metal conductors have other chemistry problems. If the entire path is bonded continuously, repair and testing may require physically severing material. Removing conductively mated dies, chips or modules for repair/replacement may entail cutting mechanical linkages and removing solder, metal-metal bonds, pins from sockets, or the like. Such breaks introduce metal fatigue (i.e. increase the likelihood of future failure modes) and contamination by conductive dust. The conductor-conductor junctions will later need to be realigned and restored. Intense localized thermal stresses from (de)soldering may also be involved. Methods for replacement pose severe constraints on the design and manufacture of conductively-coupled components in electronic systems.
Even if these or other approaches were practical in a manufacturing sense, all suffer the series inductance performance costs of conductive signalling. The performance improvement achievable with solder-bumped or wire-bonded dies is fundamentally limited by the excessive series inductance of the solder bump or wire-bond interconnection. State-of-the-art MCM technologies endure much the same performance limitation as current surface mounted IC technologies. Both families of technology require bonding wires, solder bumps, TAB, or their equivalent, to couple signals conductively between the dies and the substrate (for MCMs) or Level 2 package (for surface mounted ICs), which impose parasitic series inductance unavoidably. Parasitic inductance of 50-1000 pH is typical, and introduces significant latency, frequency limitations, and power requirements for signalling off-chip. While the best currently available MCM technologies reduce the disparity between on-die and off-die communication compared to the conventionally quoted ten-to-one ratio for surface mounted IC technologies, there remain significant penalties associated with inter-die signalling on conventional MCMs. Current MCM technologies suffer from mechanical and thermal problems, design and fabrication limitations, power costs, complexity, and expense.
With the present state of the art, technology for conductive signalling is itself directly responsible for many problems with current MCMs. It imposes mechanical and cost limitations on the density, number, and arrangement of signal leads attached to a MCM package or constituent dies in carriers, and generally requires further expense and yet more volume devoted to mounting and interconnecting the leads from chip packages onto Level 2 circuit boards or multichip carrier modules and packaging at all higher levels. Typical Level 1 packages holding a die and its conductive leads are generally much larger than the active area of the contained Level 0 die, due to requirements for spacing out contact pads and attached leads, and MCMs similarly inflate real estate requirements. On each die, fan-out of conductive leads (in order to simplify testing or wiring contact pads) takes up chip real estate. The practical need for a sufficient number of conductive pads, adequately large and well-separated, typically accounts for several tens of percent of the chip's real estate, but can range from a fraction of a percent to essentially all of the chip's real estate. With some MCM technologies pad bonding locations cannot overlay active circuits, since applied contact probe pressure or other process steps risk damaging the circuits, so pads occupy real estate at the expense of useful electronic devices. That wasted real estate costs money (about $10-1000/square inch). Since present approaches emphasize placing the pads at the perimeter of the die's active area to minimize wiring lengths (or at the center, to minimize thermal expansion mismatch), and pad count is often limited by the available perimeter bonding density of the die, many designs are "pad limited," wasting chip real estate, and/or "pin-limited," leading to large, expensive packages. Other leads which can be used as pin-substitutes include solder, eutectic, or soft metal (e.g., gold) bumps, conductive wires (e.g., copper, silver, or aluminum) lithographed on polyimide or tape, and sharpened probe needles.
Current MCM technologies require special circuitry for interconnecting dies operating at different voltage levels. Dies fabricated from different materials generally use different voltage levels, such as CMOS versus GaAs. Dies designed with different circuit technologies, even in the same material, generally use different voltage levels, such as Si TTL versus Si ECL. Even dies fabricated from the same technology may still use different voltage levels, such as CMOS in silicon at 5.0 volts versus 3.3 volts.
The circuitry supporting signalling is sensitive to manufacturing variations during wafer fabrication. Manufacturing processes produce variation in basic physical parameters (e.g., impedance, capacitance, inductance) absolutely and spatially, across a wafer or between wafers. This variation causes circuitry to produce waveforms of differing spectra, affecting shape and skew, and differing amplitude, affecting thresholds, noise susceptibility, power requirements and termination characteristics. Unreliable waveform spectra and amplitude means that the yield of dies follows a bell curve distribution, with very few able to operate at higher frequencies. Designers have therefore adopted conservative design rules to compensate for performance ranges.
Power must still be provided to the electronic devices in a system, whether the devices are packaged in MCMs or any other level. Nearly all leading MCM designs treat power leads and signal leads the same way at Level 1, which wastes space and loses important opportunities for increasing the density of signal lines.
In short, there remains a significant need for an improved method and apparatus for coupling signals among modules in a modular electronic system which mitigates one or more of the above-explained problems with the present MCM technology. While MCM technology represents an important application domain, the invention, as defined in the succeeding sections, applies broadly to all levels of electronic packaging and interconnection, such as die, wafer, board, MCM, system, etc. It is therefore the inventors' intent that the invention not be viewed narrowly, or only in the context of the preferred MCM embodiment, except where the context inarguably indicates otherwise.